Method for tracking a target in an image sequence, taking the dynamics of the target into consideration

Abstract

A method for tracking a target in a sequence of images comprises: a step of detecting objects, a temporal association step, aiming to associate the objects detected in the current image with the objects detected in the previous image based on their respective positions, a step of determining a second target in the current image according to a search area determined for a previous image of the sequence, a step of determining the detected object that best corresponds to a dynamic of a final target, a step of updating the search area for the current image based on the position of the target that best corresponds to the dynamics of the final target, and a step of searching for the final target in the search area for the current image by comparing areas of the current image with a reference model representative of the final target.

The invention claimed is: 
     
       1. A method of tracking a target in a sequence of images representing a scene, the method comprising, for each current image:
 detecting objects in the scene, each object detected being characterized by a position and a dimension in the current image, 
 a temporal association, in which an attempt is made to associate each object detected in the current image with an object detected in a previous image of the sequence as claimed in their respective positions in the images, each set of objects detected in a continuous sequence of images forming a tracklet, one of the tracklets being assigned to a first target, 
 determining a second target in the current image, the detected object best corresponding to a given search area determined for a previous image of the sequence being assigned to the second target, 
 comparing between an identity of the detected object assigned to the first target and an identity of the detected object assigned to the second target, 
 if different identities, determining the detected object that best corresponds to a dynamic of a final target, 
 if the first target best corresponds to the dynamics of the final target, updating the search area for the current image based on a position of the first target, 
 if the second target best corresponds to the dynamics of the final target, updating the search area for the current image based on a position of the second target, and 
 searching for the final target in the search area for the current image by comparing areas of the current image with a reference model representative of the final target. 
 
     
     
       2. The method of  claim 1 , further comprising, in the case where the detected object assigned to the first target is the same as that assigned to the second target, updating the search area for the current image based on the position of the first and the second target. 
     
     
       3. The method of  claim 1 , further comprising, following a temporal association extending the tracklet of the first target when said tracklet was not supplemented by an object detected in the current image, the tracklet of the first target being supplemented by a virtual observation as claimed in the dynamics of a final target. 
     
     
       4. The method of  claim 1 , comprising, in the case where no detected object has been assigned to the second target:
 determining if the object detected in the current image associated with the first target is consistent with the dynamics of the final target, and 
 in case of compatibility, updating the search area for the current image based on the position of the first target. 
 
     
     
       5. The method of  claim 1 , comprising, in the case where no detected object has been assigned to the first target:
 determining if the detected object assigned to the second target does not belong to a tracklet other than that of the target, and 
 in the case where this detected object does not belong to a tracklet other than that of the target, updating the search area for the current image based on the position of the second target. 
 
     
     
       6. The method of  claim 1 , wherein updating the search area based on the position of the second target further comprises a correction of the tracklet of the first target, the detected object assigned to the first target being substituted by the detected object assigned to the second target. 
     
     
       7. A video surveillance system, comprising a processor and a memory configured to implement a method of tracking a target in a sequence of images representing a scene, the method comprising, for each current image:
 detecting objects in the scene, each object detected being characterized by a position and a dimension in the current image, 
 a temporal association, in which an attempt is made to associate each object detected in the current image with an object detected in a previous image of the sequence as claimed in their respective positions in the images, each set of objects detected in a continuous sequence of images forming a tracklet, one of the tracklets being assigned to a first target, 
 determining a second target in the current image, the detected object best corresponding to a given search area determined for a previous image of the sequence being assigned to the second target, 
 comparing between an identity of the detected object assigned to the first target and an identity of the detected object assigned to the second target, 
 if different identities, determining the detected object that best corresponds to a dynamic of a final target, 
 if the first target best corresponds to the dynamics of the final target, updating the search area for the current image based on a position of the first target, 
 if the second target best corresponds to the dynamics of the final target, updating the search area for the current image based on a position of the second target, and 
 searching for the final target in the search area for the current image by comparing areas of the current image with a reference model representative of the final target.